Laurence C. Baker, PhD

 To assess whether the characteristics and capabilities of individual practices intending to join the early Medicare accountable care organization (ACO) programs differed from those of practices not intending to join.

 Data from a 2012-2013 national survey of 1398 physician practices were linked to 2012 Medicare beneficiary claims data to examine differences between practices intending to join a Medicare ACO and practices not intending to join a Medicare ACO.

 Differences were examined with regard to patient sociodemographic characteristics and disease burden, practice characteristics and capabilities, and cost and quality measures. Logistic regression was used to examine the differences.

 Practices intending to join were more likely to have better care management capabilities (odds ratio [OR], 1.72; 

 <.003), health information technology functionality (OR, 1.87; 

 <.001), and use of quality improvement methods (OR, 1.52; 

 <.04). They were also more likely to have had prior pay-for-performance experience (OR, 1.59; 

 <.02) and less likely to be physician-owned (OR, 0.51; 

 <.001). However, the practices with the greater capabilities still used half or less of them.

 Physician practices that intended to join the early ACO programs had greater capabilities and experience to manage risk than those practices that decided not to join. The early ACO programs thus attracted the more capable physician practices, but those practices still fell short of implementing key recommended behaviors. The findings have implications for future physician practice selection into ACOs.

Practices intending to join accountable care organizations (ACOs) in 2012 had greater capabilities than those not intending to join but implemented half or less of the capabilities. This may partially help explain the great variation found in ACO performance. The findings suggest the need for more careful targeting of practices in the future and greater investment in developing the capabilities of future practices to succeed in a value-based payment environment.

Those practices intending to join were also more likely to be hospital-owned and have prior experience with pay-for-performance payment models.

There were no differences in practice size or related variables.

There were no differences in patient demographic characteristics or illness severity/disease burden or in baseline Medicare spending per beneficiary, ambulatory care—sensitive admission rates, or 30-day unplanned hospital readmissions.

A major response to the continued growth in healthcare costs and the highly varying quality of care in the United States has been the formation of accountable care organizations (ACOs). As of 2018, there are 1011 ACOs, with at least 1 located in every state, serving approximately 32 million people.

 Enrollment has been projected to grow to 68 million by 2020.

 Despite high hopes for savings, early research suggests only modest savings, albeit with no diminution of quality and, in fact, with increases on some quality measures.

 Substantial variation in the performance of ACOs has also been noted.

There are a number of possible reasons that ACOs thus far appear to have achieved only modest savings. This paper’s contribution lies in using a unique national survey of physician practices to explore the possibility that there were differences in characteristics and capabilities between physician practices that intended to join the Medicare ACO programs in 2012 and those that did not. Presumably, those practices intending to join might have been larger, had greater access to resources, and believed they had greater ability to contain costs, achieve savings, and improve quality by doing a better job of managing patient care.

Taking a look back to the beginning of the ACO program is important for considering what may happen with the introduction of alternative payment models (APMs) and the Merit-based Incentive Payment System (MIPS) authorized to take effect in 2019 as part of the Medicare Access and CHIP Reauthorization Act. Practice leaders will need to assess their ability to take on financial risk and meet the challenge of performance metrics. Due to the importance of health information technology (HIT), care management processes (CMPs), and quality improvement (QI), we assess whether practices with strong capabilities in these areas were more or less likely to intend to join.

 Such capabilities are markers of a practice’s ability to redesign care to meet quality goals and achieve shared savings. If practices without strong capabilities were the most likely to join, then the ability of ACOs to implement the necessary changes to drive higher-value care may have been limited, with implications for those that have joined more recently and those intending to join in the future.

We draw on a unique national survey of individual physician practices in 2012-2013 to identify those choosing to join the Medicare ACO programs versus those not choosing to join. We collected data on practice ownership, size, care management capabilities, HIT use, QI processes, and related variables. We linked these data to the Medicare claims files to examine baseline differences across practices in Medicare patient characteristics and illness severity/disease burden, focusing on high-cost/high-need patients’ spending, ambulatory care—sensitive admissions (ACSAs), and 30-day unplanned hospital readmissions.

Specifically, we used the IMS Healthcare Organizational Services Database to select a national sample of physician practices, with oversampling in 17 communities that participated in the Robert Wood Johnson Foundation Aligning Forces for Quality program. Veterans Health Administration practices and academic medical center practices were excluded. We focused on practices most likely to provide care for patients with chronic illnesses, including asthma, congestive heart failure, depression, and diabetes. These included general internal medicine, family medicine, general practice, cardiology, endocrinology, and pulmonology practices.

Between January 2012 and November 2013, the physician or administrative practice leader completed a 40-minute telephone survey administered by RTI International. The survey instrument (

]) was based on prior National Survey of Physician Organizations (NSPO) surveys. A total of 1398 practices responded, yielding an adjusted response rate of 50%.

 We restricted our analyses to the practices that included at least 15% primary care physicians (PCPs), which totaled 1040 practices. Of these, 235 (22.6%) practices intended to join a CMS ACO program. These data were then linked to the 2012 Medicare beneficiary claims data using the CMS method for assigning beneficiaries to ACOs (

). A total of 868,213 beneficiaries were linked to surveyed practices; 311,116 beneficiaries were attributed to the physician practices intending to join the CMS ACO program. Beneficiaries were eligible if they were 65 years or older as of January 1, 2011; were not in the end-stage renal disease program; and were alive and enrolled in Medicare parts A and B throughout 2011 and 2012. The research protocol was approved by the Institutional Review Board of the University of California, Berkeley.

The dichotomous dependent variable was the practice’s intention to join an ACO in 2012. Specifically, respondents were asked if their organization applied to become an ACO during 2012. For our analysis, we constructed a 0/1 variable indicating those that responded affirmatively to this question.

Our principal independent variables of interest were measures of CMPs, HIT functionality, and QI activities. Each practice was measured on a 20-item care management capability index, which included items involving nurse care management, reminders and education for patients, provision of quality data to physicians, and use of a registry to identify patients with chronic illnesses. Each practice was also measured on a 14-item HIT functionality index that included various uses of an electronic health record, such as clinical decision support, collection of quality data, and electronic connectivity with patients. QI activities were measured based on whether the practice used 1 or more of 5 specific QI processes, including Plan-Do-Study-Act cycles, lean, Six Sigma, and related approaches. The elements of these 3 indices along with their reliability are shown in 

. We recoded these variables and placed the practices into dichotomous categories based on whether or not they scored in the top quartile of each composite index.

In addition, we included a number of other measures of practice characteristics in our analysis. Ownership was measured by whether the practice was physician-owned, hospital-owned, or a federally qualified health center (FQHC) or community health center (CHC). Practice size was measured by the 6 categories shown in 

 and ranged from 1 or 2 physicians to more than 100. We also included each practice’s percentage of PCPs as a continuous variable and whether or not the practice had previous experience with pay-for-performance and/or public reporting programs.

Using the 2012 Master Beneficiary Summary File Cost and Use segment, we measured total spending per beneficiary, defined as the allowed amounts paid for services by Medicare, coinsurers, and the beneficiary. We also measured spending in 4 subcategories: hospital services, physician services, postacute care, and other services. Outliers were reduced to the amount spent at the 99th percentile for each subcategory. Each subcategory was then summed to measure total spending. We standardized spending to adjust for geographic differences in Medicare payments at the county level by multiplying spending in each category by the ratio of Medicare county-level total standardized spending to total actual spending in that county in 2012.

 We used this to construct practice-level measures of mean spending.

To assess severity of illness/disease burden, we used 2011 Medicare claims and modified a method developed by Jha et al by placing each beneficiary into 1 of 5 categories of need, ranging from those predicted to have the highest need (beneficiaries with 2 or more frail conditions) to the lowest need (beneficiaries with no chronic conditions), based on total spending on care in 2012.

 For purposes of this analysis, we collapsed the category with 3 or more chronic conditions and the category with 2 or more frailty conditions into 1 category labeled high-need, and those with 2 or fewer major conditions, only minor conditions, or no other conditions into the low-need category.

We used 2 utilization-based measures of quality. The first was the number of unplanned 30-day hospital readmissions for each beneficiary, measured using the hospitalwide (all-condition) 30-day risk-standardized readmission measure, which excludes planned readmissions.

 We computed whether or not the practice was above or below the median on this measure. Second, we measured the number of ACSAs for each beneficiary for conditions, such as congestive heart failure, chronic obstructive pulmonary disease, and short-term complications of diabetes, for which good outpatient care may reduce the rate of hospital admissions.

 We computed whether or not the practice was above or below the median on this measure.

From the Medicare Master Beneficiary Summary File, we included beneficiaries’ age, sex, racial/ethnic group, whether the beneficiary was dual-eligible (covered by both Medicare and Medicaid), and whether disability was the original reason for Medicare eligibility. We used these data to calculate the share of each practice’s Medicare beneficiaries in each of the above areas. We also calculated the percentage of the practice’s patients that were in the high-need/high-complexity category versus the low-need/low-complexity category. To account for possible heterogeneity within the need and complexity categories, we also included variables for each beneficiary’s number of major and minor conditions. We also took into account the percentage of the practice’s revenue from Medicaid, measured as a continuous variable. Finally, we included the mean annual spending per Medicare beneficiary for the hospital referral region in which the beneficiary resided, in order to account for the possible influence of local practice norms.

We used a logistic model to evaluate our binary dependent variable of whether or not the practice chose to join an ACO. All analyses were conducted accessing the Medicare Virtual Data Research Center using SAS Enterprise Guide 7.1 (SAS Institute; Cary, North Carolina).

Table 1 shows that the practices that chose to join a CMS ACO program had a greater number of Medicare beneficiaries, had twice as many practices with greater than 100 physicians, were more likely to be hospital-owned, had more experience with pay-for-performance and public reporting programs, and had greater capabilities in CMPs, HIT functionality, and QI activities than practices that did not join. It is important to note that, as Table 1 illustrates, even the practices scoring in the top quartile of the QI index and the CMP index were using less than half of the available recommended capabilities and, in regard to HIT functionality, were using only half. The 2 groups did not differ on any of the characteristics of patients seen, including the percentage of high-need/high-complexity patients. There were also no differences in total spending (costs), ACSA rates, or 30-day unplanned hospital readmissions.

 indicate that the practices that joined ACOs had more prior experience with pay-for-performance programs (odds ratio [OR], 1.59) and greater capabilities in CMPs (OR, 1.72), HIT functionality (OR, 1.87), and QI (OR, 1.52). The results also indicate that the physician-owned practices and the FQHC/CHC

sites were less likely to join than the hospital-owned practices (ORs, 0.51 and 0.31, respectively). There were no differences in practice size, types of patients treated, spending, ACSAs, or 30-day unplanned readmissions. The latter 2 variables were also assessed by whether or not a practice/group had any ACSAs or not or any 30-day unplanned readmissions or not, with virtually identical results.

We expected that practices intending to join ACOs would have greater CMP, HIT, and QI capabilities in order to better manage risk and achieve shared savings. The data support this expectation. But at the same time, even those scoring in the top quartile had only between 45% and 50% of the identified capabilities in CMPs, HIT, and QI. The practices joining ACOs also had greater prior experience with private-sector pay-for-performance programs containing some of the same performance metrics and incentives as those associated with the Medicare ACO programs. But although such practices may have more such capabilities and prior experience with value-based payment programs,

 there is mixed evidence to suggest that they are yet associated with better performance.

 It is important to recognize that most such practices have been operating in a largely risk-free fee-for-service world in which the capabilities and infrastructure to provide care more efficiently and effectively have not been needed. Given the results presented, it is likely that most practices lack the capabilities needed to succeed under new risk-based payment models. Those practices already doing well, or that at least believe that they have the QI, HIT, and CMP capabilities to do so, have the opportunity to get paid for continuing to do well. Those practices that have fewer resources to invest in such capabilities are likely to fall further behind, potentially exacerbating possible disparities in patient care. Thus, “catch-up” policies and practices may be needed, including technical assistance programs, such as the CMS Transforming Physician Practice Initiative; consolidation of practices or ownership by hospitals to create or gain access to greater resources; and partnerships with insurers and nonprovider organizations that provide management and infrastructure support services.

 Recent research suggests that physicians themselves participating in Medicare ACOs are only moderately convinced that ACOs are an effective model for delivering cost-effective care,

 despite some evidence that physician-owned practices are more financially successful.

 Those practices entering into ACO arrangements today and in the future may require an even stronger set of capabilities and face an even steeper learning curve to be successful under the newly evolving value-based payment arrangements.

Although there were small differences between practices that responded to the NSPO survey versus those that did not,

 it is possible that unmeasured differences existed between the 2 groups that could result in biased estimates of the variables associated with those practices that intended to join an ACO versus those that did not. We also did not capture all practices in 2012 that intended to apply to become a Medicare ACO. Further, although we have key informant survey measures of each practice’s use of CMPs, HIT functionality, and QI processes, we do not have fine-grained measures of the quality or consistency of implementation of these processes, which are better captured through qualitative methods. Finally, we did not have access to the follow-up performance measures of the individual physician practices that intended to join an ACO in 2012, but rely on the literature cited above

 indicating that overall ACO accomplishments to date have been modest.

Physician practices intending to join an ACO in 2012 at the beginning of the program had greater care management capabilities, HIT functionality, and use of QI methods; were more likely to be hospital-owned; and had more experience with value-based payment programs than those practices that were not intending to join. There were no net differences in size, types of patients seen, or other characteristics. Thus, the early-joining physician practices did have greater capabilities to manage risk and succeed under the new value-based payment model. However, despite the involvement of such practices overall, ACO performance has been modest and, to some extent, disappointing. The finding that even the more capable practices used only 50% or less of recommended care management, HIT, and QI processes may be a reason for this relative lack of success. Forthcoming APMs and MIPS value-based payment models will likely present challenges to not only more advanced physician practices but also, in particular, to those currently lacking the needed capabilities to succeed.

The research upon which this paper is based was supported by the Commonwealth Fund under grant #20150261. The authors thank David Muhlestein, PhD, JD, of Leavitt Partners for supplying information helpful to the analysis.

 Center for Healthcare Organizational and Innovation Research (CHOIR), School of Public Health, University of California, Berkeley (SMS, PPR), Berkeley, CA; Department of Health Research and Policy, Stanford University School of Medicine (LCB), Stanford, CA; Department of Economics, Georgia State University (MFP), Atlanta, GA; Department of Healthcare Policy and Research, Weill Cornell Medicine (LPC), New York, NY.

 Dr Shortell is a member of the Advisory Board of the Centene Corporation. Dr Casalino is a board member of the Hospital Research and Educational Trust of the American Hospital Association and is part of the American Medical Association Committee on Professional Satisfaction. The remaining authors report no relationship or financial interest with any entity that would pose a conflict of interest with the subject matter of this article.

 Concept and design (SMS, PPR, LCB, MFP, LPC); acquisition of data (SMS, PPR, LPC); analysis and interpretation of data (SMS, PPR, LCB, MFP, LPC); drafting of the manuscript (SMS, PPR); critical revision of the manuscript for important intellectual content (SMS, LCB, MFP, LPC); statistical analysis (PPR, MFP); obtaining funding (SMS); administrative, technical, or logistic support (PPR); and supervision (SMS).

 Stephen M. Shortell, PhD, MPH, MBA, University of California, Berkeley School of Public Health, 2121 Berkeley Way, Room 5317, Berkeley, CA 94720. Email: shortell@berkeley.edu.

1. Muhlestein D, Saunders R, Richards R, McClellan M. Recent progress in the value journey: growth of ACOs and value-based payment models in 2018. 

 Blog website. healthaffairs.org/do/10.1377/hblog20180810.481968/full. Published August 4, 2018. Accessed September 5, 2018.

2. Ahier B. The ACO journey continues. LinkedIn website. linkedin.com/pulse/aco-journey-continues-brian-ahier. Published September 16, 2016. Published July 6, 2017.

3. McWilliams JM, Chernew ME, Landon BE, Schwartz AL. Performance differences in year 1 of Pioneer accountable care organizations. 

. 2015;372(20):1927-1936. doi: 10.1056/NEJMsa1414929.

4. McWilliams JM, Hatfield LA, Chernew ME, Landon BE, Schwartz AL. Early performance of accountable care organizations in Medicare. 

. 2016;374(24):2357-2366. doi: 10.1056/NEJMsa1600142.

5. Colla CH, Lewis VA, Kao LS, O’Malley AJ, Chang CH, Fisher ES. Association between Medicare accountable care organization implementation and spending among clinically vulnerable beneficiaries [erratum in 

. 2017;177(11):1702-1703. doi: 10.1001/jamainternmed.2017.4130]. 

. 2016;176(8):1167-1175. doi: 10.1001/jamainternmed.2016.2827.

Articles

Physician practices intending to join Medicare accountable care organizations (ACOs) in 2012 had greater capabilities in health information technology, care management processes, and quality improvement methods than those not intending to join, but they still were far from using all recommended behaviors to manage risk.

The Characteristics of Physician Practices Joining the Early ACOs: Looking Back to Look Forward

To examine the association between annual premiums for health plans available in Federally Facilitated Marketplaces (FFMs) and the extent of competition and integration among physicians and hospitals, as well as the number of insurers.

We used observational data from the Center for Consumer Information and Insurance Oversight on the annual premiums and other characteristics of plans, matched to measures of physician, hospital, and insurer market competitiveness and other characteristics of 411 rating areas in the 37 FFMs.

estimated multivariate models of the relationship between annual premiums and Herfindahl-Hirschman indices of hospitals and physician practices, controlling for the number of insurers, the extent of physician—hospital integration, and other plan and rating area characteristics.

 Premiums for Marketplace plans were higher in rating areas in which physician, hospital, and insurance markets were less competitive. An increase from the 10th to the 90th percentile of physician concentration and hospital concentration was associated with increases of $393 and $189, respectively, in annual premiums for the Silver plan with the second lowest cost. A similar increase in the number of insurers was associated with a $421 decrease in premiums. Physician—hospital integration was not significantly associated with premiums.

 Premiums for FFM plans were higher in markets with greater concentrations of hospitals and physicians but fewer insurers. Higher premiums make health insurance less affordable for people purchasing unsubsidized coverage and raise the cost of Marketplace premium tax credits to the government.

Do insurers pass on higher prices to health plan customers for the services of more consolidated providers? There is little empirical evidence on this issue in the context of the Health Insurance Marketplaces established by the Affordable Care Act (ACA). We provide evidence consistent with the idea that provider market power may be an important driver of plan costs in ACA markets. Moreover, our results suggest that the role of provider market power in the price of health insurance may be more important than the degree of competition among insurance companies, which has received more attention.

ACA health plans were more expensive in areas with concentrated providers.

Provider concentration may be as predictive of health plan prices as insurer competition.

Premiums for health insurance offered through the Affordable Care Act (ACA)’s Health Insurance Marketplaces (HIMs) vary substantially across the country.

 Differences in health insurance premiums throughout geographic areas affect not only the affordability of coverage for individuals, but also the cost of ACA coverage to the government because government subsidies for relatively low-income families are linked to premiums.

Researchers have documented that premiums are lower in marketplaces with greater competition among insurers.

 The effect of competition among healthcare providers on premiums for Marketplace plans, however, has received less attention. Increasing consolidation among and integration between physicians and hospitals have led to higher prices for physician and hospital services.

 If insurers pass on these higher prices to consumers in the form of higher premiums, greater concentration in provider markets could lead to higher premiums in the Marketplaces. Although a recent case study of Marketplaces in New York and California provided preliminary evidence that hospital competition may be related to Marketplace premiums,

 there is no systematic evidence on whether differences in the structure of provider markets contribute to geographic variances nationwide in the premiums for Marketplace plans. This market is an important new setting in which to examine the relationship between provider concentration and plan premiums due to the relatively widespread offering of narrow network plans.

In this paper, we examined whether the concentration of local hospital and physician markets, the degree of physician—hospital integration, and the number of insurers were related to the premiums for 2015 plans sold on the Federally Facilitated Marketplaces (FFMs).

We used publicly available data from the Center for Consumer Information and Insurance Oversight (CCIIO) on annual premiums for each plan available in 2015 on the FFMs.

 All plans on the Marketplaces are classified according to their metal level (Bronze, Silver, Gold, or Platinum), which corresponds to the actuarial value of the plan. We focused our analysis on the premiums of the second-lowest-cost Silver plan (SLCSP) and of the overall lowest-cost plan (LCP) of any metal level (excluding Catastrophic plans) in each rating area. Within a demographic group, insurers must charge the same premium for a given plan within a rating area. Rating areas are typically a collection of counties defined by states. We included all 411 rating areas for the FFMs in 37 states. The premiums in our sample did not include any means-tested subsidies or tax credits.

The SLCSP, the benchmark for setting federal subsidies for insurance purchases, is likely to be a focus of many consumers. The LCP is also likely to attract many consumers; research on the choice of health insurance plans shows that individuals tend to place too much emphasis on the role of premiums relative to other plan characteristics and disproportionately enroll in the lowest-premium plans.

 Although plans may vary premiums by enrollee age, family structure, and smoking status, we limited our analysis to 1 rate category for each plan: a 50-year-old nonsmoker buying individual, rather than family, coverage.

Measures of Provider Competition and Vertical Integration

For each rating area, we developed measures of hospital and physician competition and the degree of vertical integration between physicians and hospitals. Following previous work, we computed a Herfindahl-Hirschman Index (HHI) for hospitals and physicians.

 The HHI is a standard measure used by the Federal Trade Commission (FTC) and Department of Justice (DOJ) to assess competition.

 HHIs range from near 0 to 1. They are low in markets served by many providers, signaling a more competitive market, and reach the maximum of 1 in a monopoly market served by a single provider.

To develop a rating area—level measure of hospital HHI, we first calculated for each hospital an admission-weighted average of the HHIs of the patient zip codes that it serves, based on Medicare claims and enrollment data for a 100% sample of traditional Medicare beneficiaries in 2011. Using the same data, we then identified the set of hospitals used by patients in each rating area and computed an admission-weighted average of the HHIs of the hospitals serving a rating area. Our hospital competition measures did not, at any point in their construction, assume that patients residing in a given rating area use only hospitals in that rating area. We accounted for hospital system structures when constructing the HHI measure by assuming that hospitals within a system bargain jointly; in addition, we controlled for the differences in the prevalence of hospital systems across rating areas in the regression specifications.

We used an analogous approach for creating measures of physician competition.

 We defined physician practices as sets of physicians reporting the same specialty who billed Medicare under the same tax ID, meaning that they are part of the same financially integrated organization.

 For each practice—specialty combination, we calculated a Medicare payment–weighted average of the HHIs of the patient zip codes it serves based on patient flows observed in Medicare data. To derive a single physician HHI at the rating area level, we computed the Medicare payment–weighted average of the specialty-specific HHIs of all practices used by patients in each rating area.

We constructed a rating area—level measure of vertical integration between physicians and hospitals using data from the American Hospital Association’s 2011 survey of hospitals, in which hospitals reported information about their relationships with physicians.

 We identified hospitals that reported participating in fully integrated physician organizations, closed physician—hospital organizations, open physician–hospital organizations, and independent practice associations. For each hospital, we constructed an admission-weighted average of the patient zip code–level market shares held by hospitals of each type using the Medicare claims and enrollment data described above. Using these same data, we then computed for each rating area an admission-weighted average of the density of each type of vertical integration facing each hospital serving the area. For analysis, we summed the shares of the 4 types of integration to construct 1 summary measure of the prevalence of vertical integration in the markets of hospitals used by patients residing in each rating area. This measure varied from 0% to 100%, increasing with the prevalence of patients using hospitals participating in vertically integrated arrangements. Greater detail on each measure is available in other studies.

To examine the relationship between Marketplace premiums and provider market power, we estimated cross-sectional regressions in which each observation was a plan/rating area combination. The dependent variable was the premium for either the SLCSP or the LCP as defined above. The independent variables of interest included hospital and physician HHIs, number of insurers, and degree of vertical integration between physicians and hospitals. The models included continuous measures of these variables. We found no important differences from using less parametric specifications of market structure.

The models also included controls for other characteristics of the health plans and rating areas. The health plan controls included measures of plan type, cost sharing, provider coverage, and offering of chronic condition management.

] includes a complete list of these variables.) We used Medicare claims and the American Hospital Association 2011 survey of hospitals to construct other rating area measures of hospital market characteristics in the same way that we constructed rating area densities of vertical integration. We used county-level Area Resource File data to calculate rating area population characteristics and controlled for variations in practice costs using the Medicare geographic practice cost index. Finally, we used the information about FFM health insurance plans released by CCIIO to calculate the number of insurance companies and issuers (see eAppendix for details of how insurance company was defined) competing in the Marketplaces in each rating area. The models also included state indicators in order to control for other characteristics of the states, such as the regulatory climate, insurance market features, and any state-specific provider and population characteristics.

We used the estimated coefficients from the multivariate models to compute the predicted changes in premiums that would result from moving from the 10th to the 90th percentile of hospital and physician HHI, the number of insurers, and the extent of vertical integration. For statistical inference, we used standard errors clustered at the plan level to allow for unobserved differences within plans across rating areas. The eAppendix reports several robustness checks, including results from models that include the average premium of all plans offered in each rating area, metropolitan and nonmetropolitan rating areas, and Marketplaces from all states, rather than only FFMs, but with fewer controls for plan characteristics. The Stanford University Review Board reviewed the study protocol and granted a waiver of consent.

The average annual premium (for a nonsmoking individual aged 50 years) across all 4580 plans in the dataset was $5378 (SD = $1443). The average premium was $4718 (SD = $784) for the SLCSP and $3651 (SD = $656) for the LCP.

The markets for both hospital and physician services were, on average, relatively highly concentrated in the geographic areas that we studied. The average hospital HHI across the 411 rating areas was 0.56 (SD = 0.14), and the average physician practice HHI was 0.41 (SD = 0.09). The FTC and DOJ typically consider markets with HHIs above 0.25 to exhibit a high degree of concentration. The average of our vertical integration measure was 0.56 (SD = 0.28), indicating that, on average, 56% of patients used hospitals that participated in vertically integrated arrangements.

Although relatively high on average, the extent of provider market concentration varied across rating areas (

). The hospital HHI averaged 0.38 at the 10th percentile and 0.70 at the 90th percentile rating area. The physician HHI measure was slightly less variable: 0.27 and 0.50, respectively. The extent of vertical integration was highly variable across markets, ranging from 0.06 at the 10th percentile to 0.89 at the 90th percentile.

Premiums for exchange plans were higher in rating areas with more concentrated provider markets (Table). The average annual premium for the SLCSP was $4330 in the rating areas with the least concentrated hospital markets (10th percentile), $5014 in areas with the most concentrated hospital markets (90th percentile), $4461 in the rating areas with the least concentrated physician markets, and $4873 in areas with the most concentrated physician markets. The patterns were similar for the relationship between premiums for the LCP and physician and hospital concentration. These differences are highly statistically significant. In unadjusted analyses, vertical integration was not statistically significantly associated with premiums.

Premiums for exchange plans and provider market concentration were positively correlated even after adjusting for an extensive set of plan- and market-level control variables. 

 present the predicted change in premiums from moving from the 10th to the 90th percentile of the respective market power measures, based on the coefficient estimates from the regression. A change from the 10th to the 90th percentile in the physician HHI was associated with a $393 increase in the annual premium for the SLCSP and a $386 increase in the annual premium for the LCP (

<.001). These changes were also economically significant, corresponding to about an 8% increase relative to the average premium for the SLCSP and a 10% increase relative to the average premium for the LCP. Hospital market power had a similar association with premiums, although the magnitude was approximately half as large. We did not find a statistically or economically significant relationship between premiums and the extent of vertical integration.

To provide context for the estimated premium changes associated with provider concentration, Figures 1 and 2 also report the adjusted association between the number of insurers participating in the rating area and premiums. Our point estimates were quite close to the estimates of insurer effects that have been reported in the previous literature.

 A change from the 10th to 90th percentile in the number of insurers participating in the rating area was associated with a $421 decrease in the annual premium for the SLCSP (

 <.01) and a $449 decrease in the annual premium for the LCP (

 <.001). These associations for insurers were comparable to the associations we observed for provider concentration. The association we observed for physician market concentration, in particular, is nearly as large as that for the number of insurers participating in the market.

Our study findings demonstrate that premiums for health plans in the ACA HIMs are higher in rating areas with less competition among physicians, hospitals, and insurers. These findings are consistent with research demonstrating that prices for hospital and physician services are higher in more concentrated markets. The findings suggest that insurers pass on these higher prices for healthcare services to consumers in the form of higher premiums for coverage.

Our results provide 1 potential explanation for the prevalence of narrow network plans on the insurance Marketplaces: the use of selective contracting to limit the impact of provider concentration on premiums by avoiding providers with the most market power. Our results suggest, however, that even if narrow networks were set up to limit the impact of provider market power, they were not sufficient to eliminate the association between provider market structure and premiums. Regulatory requirements to cover certain types of providers and services may have limited insurers’ ability to avoid costly providers. At the same time, beneficiaries’ muted price sensitivity due to the subsidization of Marketplace premiums may have limited insurers’ incentive to avoid costly providers. It is also possible that the development of narrow networks was more effective in negotiating lower prices for hospital than for physician services in concentrated markets, explaining the difference we observed in the effect of market concentration in the 2 sectors.

Although other studies have found that prices for hospital services are higher in markets with greater levels of physician—hospital integration, we did not find evidence that health plan premiums were higher in more integrated markets. We speculate that narrow networks may have been reasonably effective at avoiding or negotiating with vertically integrated providers and would provide 1 reason for why we do not find a strong association between premiums and vertical integration.

Our study has several limitations. Our analyses are cross-sectional, and our estimates could be biased if unobserved characteristics of plans or rating areas correlated with both market competitiveness and premiums. For example, our controls did not include detailed measures of provider network breadth due to data limitations. In rating areas with higher provider concentration, insurers may be more likely to offer narrow networks, which would induce a downward pressure on premiums, biasing our estimates toward 0. Our models included an extensive set of control variables, including state fixed effects, but the possibility of bias due to omitted variables remains.

In addition, our measures of provider market structure were derived from 2011 Medicare claims data. The premiums we studied were largely set by summer 2014, leaving a lag of more than a year between the HHIs and the premiums and creating the possibility that our measures inaccurately characterized the market conditions in place when the premiums were set. Although market conditions evolve slowly, perhaps to the point that any bias from this lag would be small, we cannot rule out that measurement error from this source would cause us to underestimate the relationship between market characteristics and premiums. In particular, it is conceivable that the measurement error from the lag was more important for the vertical integration measure, which could be another reason for why our estimates of the effect of vertical integration were small and statistically insignificant.

The HHIs and vertical integration measures we used rely on patient flows observed in Medicare data. Medicare is one of the few sources of sufficiently detailed data to construct these types of measures, but these data may not represent the patient flows relevant to the non-Medicare market. We do expect that Medicare data will represent the majority of hospitals serving the under-65 market. Medicare claims also reflect care delivered by a very large share of active physicians, and the set of physicians who billed traditional Medicare should substantially overlap with the set of physicians providing services to privately insured patients. Nonetheless, this may also be a source of measurement error that could have caused us to understate the strength of the association between market characteristics and premiums.

Premiums for insurance offered in the FFMs were higher in markets with greater concentrations of hospitals and physicians and smaller numbers of insurers. Health insurance is less affordable for people purchasing unsubsidized coverage in these areas and is costlier to the government through subsidies in the form of Marketplace premium tax credits. To the extent that policy initiatives to promote coordinated care encourage the consolidation of providers or insurers, they may also have the unintended effects of making health insurance less affordable for consumers.

The authors thank Alan Jaske for outstanding research assistance.

Stanford University (MP, MKB, DPK, LCB), Stanford, CA; National Bureau of Economic Research (MP, MKB, DPK, LCB), Cambridge, MA.

 Dr Kessler has provided expert testimony for insurers and hospitals and has received lecture fees for integrated delivery system executive education. The remaining authors report no relationship or financial interest with any entity that would pose a conflict of interest with the subject matter of this article.

 Concept and design (MP, MKB, DPK, LCB); acquisition of data (MP, DPK, LCB); analysis and interpretation of data (MP, MKB, DPK, LCB); drafting of the manuscript (MP, MKB, DPK); critical revision of the manuscript for important intellectual content (MP, MKB, LCB); statistical analysis (MP); and supervision (MP).

 Maria Polyakova, PhD, Stanford University, Redwood Bldg, Rm T111, 150 Governor’s Ln, Stanford, CA 94305. Email: maria.polyakova@stanford.edu.

1. Gabel JR, Whitmore H, Call A, Green M, Oran R, Stromberg S. Modest changes in 2016 health insurance marketplace premiums and insurer participation. The Commonwealth Fund

website. commonwealthfund.org/publications/blog/2016/jan/2016-health-insurance-marketplace-premiums. Published January 27, 2016. Accessed January 10, 2018.

2. Sheingold S, Nguyen N, Chappel A. Competition and choice in the health insurance marketplaces, 2014-2015: impact on premiums. Office of the Assistant Secretary for Planning and Evaluation website. aspe.hhs.gov/basic-report/competition-and-choice-health-insurance-marketplaces-2014-2015-impact-premiums. Published July 30, 2015. Accessed January 10, 2018.

3. Burke A, Misra A, Sheingold S. Premium affordability, competition, and choice in the health insurance marketplace, 2014. Office of the Assistant Secretary for Planning and Evaluation website. aspe.hhs.gov/report/premium-affordability-competition-and-choice-health-insurance-marketplace-2014. Published June 18, 2014. Accessed January 10, 2018.

Premiums in federally facilitated Affordable Care Act Health Insurance Marketplaces are higher in geographic areas with more concentrated hospitals and physician practices and fewer insurers.

ACA Marketplace Premiums and Competition Among Hospitals and Physician Practices

To investigate the source of the weak correlation across geographic areas between Medicare and private insurance spending.

 We obtained Medicare spending data at the hospital referral region (HRR) level for 2007 from the Dartmouth Atlas, and commercial claims from large employers for 2007 from the Truven MarketScan Database. We constructed county-level data on hospital market structure from Medicare patient flows and obtained county-level data on the Medicare wage index from the Centers for Medicare & Medicaid Services website. We aggregated these sources to the HRR level. We decomposed Medicare and private spending into 2 components: price and volume. We also decomposed Medicare and private prices into 2 components: a common measure of cost and a sector-specific markup. We computed correlations between Medicare and private prices and volumes, and the correlation of each sector’s price and volume with cost and markup.

 We found that Medicare and private prices are strongly positively correlated, largely because both are keyed off of common costs. Consistent with previous work, we found that Medicare and private volumes are strongly positively correlated as well.

The weak correlation between Medicare and private spending is consistent with these 2 empirical regularities. It is mathematically due to negative correlations between each sector’s price and the other sector’s volume. In particular, we found that private prices have important spillover effects on Medicare volume. Future research on the effects of competition should take account of this phenomenon.

This retrospective, descriptive analysis investigated the source of the weak correlation across geographic areas between Medicare and private insurance spending.

		Medicare and private prices are strongly positively correlated, largely because both are keyed off of common costs. Medicare and private volumes also are strongly positively correlated.

		The weak correlation between Medicare and private spending is consistent with these 2 empirical regularities. It is mathematically due to negative correlations between each sector’s price and the other sector’s volume.

		Private prices have important spillover effects on Medicare volume. Future research on the effects of competition should consider this phenomenon.

Area variation in Medicare fee-for-service spending has been extensively documented. However, the extent to which area variation in spending by private insurers mirrors the variation in Medicare is an open empirical issue. Numerous studies have found a significant positive correlation between Medicare and the privately insured in terms of service use, such as hospital admissions, days per admission, or days in the intensive care unit at the end of life.

 In contrast, the correlation between Medicare and private spending is weakly positive or even negative, depending on the type of service and the specifics of the populations in question.

Understanding why Medicare and private spending are so weakly correlated is important. One hypothesis for the weak correlation in pending is that the prices paid by Medicare are weakly (or even inversely) correlated with the prices paid by private insurers. If true, this could be evidence that Medicare’s prices fail to reflect costs, perhaps due to political manipulation of reimbursement rates, or that private prices fail to reflect costs, perhaps due to provider market power. These explanations have very different implications for Medicare policy.

Yet no study has tested this hypothesis directly. More generally, no study has offered a unified framework in which to understand the relationships between the components of Medicare and private insurer spending. This study sought to fill this gap. We decomposed Medicare and private insurance spending into 2 components: a price index and price-adjusted spending, which we used as a measure of quantity. We calculated the simple correlations among the components of spending and the correlations between the components, the Medicare wage index, and price-cost markups.

These correlations rejected the hypothesis that Medicare’s prices are weakly or inversely correlated with those of private insurance. Rather, they are strongly positively correlated, largely reflecting differences in costs (as measured by the wage index) across areas. We provide evidence for an alternative explanation for the strong positive correlation between Medicare and private utilization and a weak correlation between Medicare and private spending.

We examined how the components of inpatient spending among fee-for-service Medicare beneficiaries are related to the components of spending among a nonelderly, privately insured population. We wrote inpatient spending per enrollee in either Medicare or private coverage from hospital referral region (HRR) i as the product of a price index for HRR i and quantity (price-adjusted spending) per enrollee in i:

	Spending per enrollee in HRR = price index

Strictly speaking, geographic price indices for enrollees in fee-for-service Medicare and private insurance plans measure different things. Medicare reimbursement rates are determined by a complex nationwide administered price system, so geographic variation in Medicare “prices” necessarily arises out of geographic variation in whatever factors the Medicare formula considers. In contrast, commercial rates depend on numerous unrelated negotiations between insurers and providers, so geographic variation in commercial prices depends on a much less well-specified, broader range of influences.

We adopted the formula used in the Dartmouth Atlas for our Medicare price index.

 According to the formula, the Medicare price level in an HRR is equal to the ratio of total spending in the HRR divided by base diagnosis-related group (DRG) payments plus adjusted outlier payments. Thus, the formula captures variation across areas in the wage index and in medical education and disproportionate-share (graduate medical education, indirect medical education, and disproportionate share hospital) payments, given each HRR’s DRG mix and outlier payments. Our private insurance price index is equal to total spending in an HRR divided by what spending would have been if all hospitals in the HRR charged the national average set of DRG prices. This index captures variation across areas in the prices that insurers pay, given each HRR’s DRG mix.

We also decomposed the price indices into 2 factors: the part that is due to differences in costs that are common to Medicare and the private sector, and a markup, which we allowed to vary across sectors. We used the Medicare wage index to represent common costs. We defined the markup as the residual from a regression of the log of the price index on the log of the Medicare wage index. Thus, for either the Medicare or private insurer price index in HRR i:

	Price index in HRR = f(Medicare wage index in HRR)

where f(.) is defined as exp[a + b*ln(wage index)], and a and b are coefficients from the regression described above. We provide a formal presentation of our decomposition in the 

We used data from several sources. First, we used HRR-level data from the Dartmouth Atlas website for 2007. These data included total Medicare inpatient spending per admission on fee-for-service beneficiaries, total price-adjusted inpatient spending per admission, the number of elderly beneficiaries, and the number of admissions per beneficiary. All of these variables are adjusted by Dartmouth for differences in the age, sex, and racial composition of areas. From this information, we constructed the price index described above. These data cover a total of 27,978,661 individuals in all 306 HRRs.

Second, we used Truven MarketScan Commercial Claims and Encounters data on the nonelderly privately insured from 2007. We limited our sample to individuals with preferred provider organization insurance and imposed the additional selection criteria suggested by Baker and colleagues.

 These data include information on the enrollees’ county of residence, which we mapped to HRRs according to the method proposed by Chernew and coauthors.

 We adjusted the variables for differences in age and sex across HRRs (we could not adjust for differences in race because the MarketScan data do not include race). These data cover a total of 14,902,153 beneficiaries in all 306 HRRs.

Third, we used data on the Medicare wage index by county from the Centers for Medicare & Medicaid Services website, which we mapped to HRRs using the method described above.

Fourth, we used data on hospital market structure that we constructed from individual-level 2007 data on Medicare patient flows according to the method proposed by Kessler and McClellan

; we provide details in the Appendix. We counted hospitals in a common system as commonly owned. 

 presents the enrollee-weighted and unweighted means and standard deviations of the variables that we analyzed. These descriptive statistics reflect the well-known properties of Medicare and private insurance area variation. There is considerable variation across areas in Medicare spending, even after adjusting for differences in prices.

 Spending per enrollee in the MarketScan data is much lower, reflecting the much lower hospital admissions rate in the nonelderly population. The enrollee-weighted means of the Medicare and MarketScan price indices are both 1 by construction.

Total spending per enrollee in MarketScan is slightly less variable across areas than total spending per enrollee in Medicare; the ratio of the standard deviation to the mean is 0.159, compared with 0.181 in Medicare. But the portion of this variation due to variation in prices is much greater in MarketScan than Medicare; the standard deviation of the MarketScan price index is 0.179 compared with 0.129 in Medicare, which is consistent with previous work.

 This is reflected in the fact that the standard deviation across areas of the price markup over the wage index is 3 times greater in MarketScan.

presents the correlation between Medicare and MarketScan total spending, quantities, and prices (

 values for the null hypothesis of zero correlation are in parentheses). Total spending is weakly positively correlated (r = 0.1135, 

 = .0473), but quantity, as measured by price-adjusted spending, is very strongly positively correlated (r = 0.6467, 

 <.0001). This finding is consistent with the previous literature, which finds strong positive correlations between the rates of use of particular services in the  elderly and nonelderly populations. It is also consistent with the hypothesis that physicians develop a single practice style that they use for patients of different ages and insurance statuses. Medicare and MarketScan prices are also strongly positively  correlated (r = 0.4476, 

 <.0001). This rejection of the hypothesis that the prices paid by Medicare are weakly (or even inversely) correlated with the prices paid by private insurers suggests that another factor is responsible for the weak correlation between  Medicare and MarketScan total spending.

Instead, according to Table 2, the weak correlation between sectors’ total spending is due to the negative correlation between each sector’s price and the other’s quantity. In other words, Medicare volume is low where private prices are high (r = —0.4324), and MarketScan volume is low where Medicare prices are high (r = –0.3748). Mathematically, this is what is causing the weak  correlation between Medicare and MarketScan total spending. (If Medicare spending is the product of prices a and quantities b, and  MarketScan spending is the product of prices c and quantities d, then Cov(ab, cd) ≈ w

) This is consistent with a wide volume of literature that finds that price reductions in Medicare induce increases in private volume.

 investigates this result further. It presents the correlations of the sectors’ prices and quantities with the Medicare wage index and each sector’s markup. According to the first column, the wage index is strongly positively correlated with both sectors’ prices, although more strongly with Medicare prices (r = 0.9001) than with private prices (r = 0.5262). It is also negatively correlated with both Medicare(r = —0.2548) and MarketScan (r = –0.4123) quantities. For example, the 3 HRRs with the lowest price-adjusted Medicare spending (Bend, Oregon, at $2644; San Luis Obispo, California, at $2677; and Santa Barbara, California, at $2763) all have relatively high wage indices (1.04, 1.12, and 1.12, respectively); by comparison, the 3 HRRs with the highest adjusted   Medicare spending (Alexandria, Louisiana, at $5666; McAllen, Texas, at $5836; and Monroe, Louisiana, at $5925) all have relatively low wage indices (0.78, 0.87, and 0.79, respectively). If we assume the wage index measures costs common to both sectors, when the wage index falls, both Medicare and private patients become absolutely more profitable, but Medicare patients become relatively less profitable than private patients, because the correlation between the wage index and the Medicare price is higher. This explains why a unit decrease in the wage index leads to increases in both Medicare and private quantities, with a greater increase in the private quantity.

According to the second column, the MarketScan markup is strongly positively correlated with the MarketScan price (r = 0.8739), and negatively correlated with both the Medicare (r = —0.3576) and the MarketScan (r = –0.2958) quantities. When the MarketScan markup rises, private quantity falls, and Medicare quantity falls even more. The 

 translates the negative correlation between MarketScan markup and Medicare quantity into dollar terms. The HRRs with the lowest MarketScan markups have average price-adjusted Medicare spending of $4426, whereas those with the highest markups have Medicare spending of $3808—a difference of almost 1 standard deviation. This is consistent with a model in which private demand slopes downward and providers respond to the declining profitability of private patients by increasing the services they deliver to Medicare patients.

According to the third column of Table 3, the Medicare markup is strongly positively correlated with the Medicare price (r = 0.4431, 

 <.0001), but roughly uncorrelated with the MarketScan price and the Medicare or MarketScan quantity. This is likely due to the minimal amount of variation in the Medicare markup; we discuss alternative explanations for this finding in the Discussion section.

Area variation in Medicare spending has long been viewed as evidence of inefficiency in the program.

 At the same time, health services researchers have observed that area variation in private insurance spending is only weakly positively (or even negatively) correlated with Medicare variations. Given that physicians likely have a single practice style that they use for both elderly and nonelderly patients, this presents a puzzle: if utilization is correlated across sectors, why isn’t spending?

In this study, we explain the source of this apparent anomaly. We decomposed Medicare and private insurance spending into 2 components: a price index and price-adjusted spending, which we use as a measure of quantity or volume. We showed that the weak correlation in overall spending is not due to weak or inverse correlation in Medicare and private prices. Not surprisingly, Medicare and private prices are strongly positively correlated across areas, largely because both are keyed off common costs (as measured by the Medicare wage index). Instead, the weak correlation is due to the negative correlation between each sector’s price index and the other’s volume.

We documented 2 channels through which this negative cross-correlation might occur. First, increases in common costs are associated with increases in both Medicare and private insurance prices, and increases in private prices are associated with decreases in private volumes. Second, increases in private insurance markups are associated with decreases in Medicare volumes. Previous research on the sources of Medicare variation has focused on beneficiary health, socioeconomic status, and preferences. Our results point to another possibility: prices in the private, under-age-65 insurance market. In analyses not reported in the tables, we showed that hospital market concentration is a likely source of this second effect: the correlation between private markups and the hospital HHI is 0.3797 (

To our knowledge, ours is the first study to identify this effect. The literature on the spillover effects of prices on utilization focuses on the impact of changes in public prices on private quantities rather than the impact of private prices on public quantities.

 We provide empirical evidence for the hypothesis proposed by Chernew and colleagues: that providers respond to the declining profitability of private patients by reducing the share of time and resources they devote to them compared with public patients.

Our analysis has significant limitations. It does not identify the causal relationships that are at the root of the negative association between one sector’s prices and the other’s volume. For example, we cannot say whether high private prices themselves cause low Medicare volumes, or whether some underlying third factor (like common costs or hospital market concentration) causes both. For purposes of policy, distinguishing between these alternatives is important. To do this would require specification of a formal model of the process through which prices and volumes are determined, which was beyond the scope of this study. In addition, our analysis was limited to inpatient acute care hospital spending. Results for outpatient services, prescription drugs, and postacute care may differ, thus complicating the explanations that we offer for the weak correlation in overall spending. Finally, our study design had only minimal independent variation in public prices. Therefore, the fact that we did not find significant spillovers from public prices to private volumes cannot be interpreted as evidence of the absence of such an effect; understanding the relative importance of private-to-public and public-to-private spillovers is an important topic for future work.

Nonetheless, our work shows that the weak correlation between spending in the 2 sectors does not, by itself, imply anything about  the processes by which prices are determined, or about the relative efficiency in one sector versus the other. However, our finding that hospital market concentration is strongly positively correlated with private payer markups supports the concern voiced by other investigators that private sector purchasers are more vulnerable to provider market power.

Future research on the policy  implications of area variations should take more careful account of this issue.  

 Frome Stanford University (LCB, KB, DPK), Stanford, CA.

 The authors report no relationship or financial interest with any entity that would pose a conflict of interest with the subject matter of this article.

 Concept and design (LCB, KB, DPK); acquisition of data (LCB, KB, DPK); analysis and interpretation of data (LCB, KB, DPK); drafting of the manuscript (LCB, KB, DPK); critical revision of the manuscript for important intellectual content (LCB, KB, DPK); statistical analysis (LCB, KB, DPK); obtaining funding (LCB, KB), administrative, technical, or logistic support (LCB, KB, DPK); and supervision (LCB, KB, DPK).

 Dr Daniel P. Kessler, Stanford University Law, GSB, and Hoover Institution, 434 Galvez Mall, Stanford, CA 94305. E-mail: fkessler@stanford.edu.1. Chernew ME, Sabik LM, Chandra A, Gibson TB, Newhouse JP. Geographic correlation between large-firm commercial spending and Medicare spending. Am J Manag Care. 2010;16(2):131-138.

2. Baker LC, Fisher ES, Wennberg JE. Variations in hospital resource use for Medicare and privately insured populations in California. Health Aff (Millwood). 2008;27(2):w123-w134.

3. Stiefel M, Feigenbaum P, Fisher ES. The Dartmouth Atlas applied to Kaiser Permanente: analysis of variation in care at the end of life. Perm J. 2008;12(1):4-9.

The weak correlation between Medicare and commercial insurance spending is due to negative correlations between each sector's price and the other sector's volume.

Laurence C. Baker, PhD

Articles

The Characteristics of Physician Practices Joining the Early ACOs: Looking Back to Look Forward

ACA Marketplace Premiums and Competition Among Hospitals and Physician Practices

Why Are Medicare and Commercial Insurance Spending Weakly Correlated?

Market-Level Health Maintenance Organization Activity and Physician Autonomy and Satisfaction